Cholesterol Contributes to Risk, Severity, and Machine Learning-Driven Diagnosis of Lyme Disease

Abstract

Background: Lyme disease is the most prevalent vector-borne disease in the US, yet its host factors are poorly understood and diagnostic tests are limited. We evaluated patients in a large health system to uncover cholesterol's role in the susceptibility, severity, and machine learning-based diagnosis of Lyme disease.

Methods: A longitudinal health system cohort comprised 1 019 175 individuals with electronic health record data and 50 329 with linked genetic data. Associations of blood cholesterol level, cholesterol genetic scores comprising common genetic variants, and burden of rare loss-of-function (LoF) variants in cholesterol metabolism genes with Lyme disease were investigated. A portable machine learning model was constructed and tested to predict Lyme disease using routine lipid and clinical measurements.

Results: There were 3832 cases of Lyme disease. Increasing cholesterol was associated with greater risk of Lyme disease and hypercholesterolemia was more prevalent in Lyme disease cases than in controls. Cholesterol genetic scores and rare LoF variants in CD36 and LDLR were associated with Lyme disease risk. Serological profiling of cases revealed parallel trajectories of rising cholesterol and immunoglobulin levels over the disease course, including marked increases in individuals with LoF variants and high cholesterol genetic scores. The machine learning model predicted Lyme disease solely using routine lipid panel, blood count, and metabolic measurements.

Conclusions: These results demonstrate the value of large-scale genetic and clinical data to reveal host factors underlying infectious disease biology, risk, and prognosis and the potential for their clinical translation to machine learning diagnostics that do not need specialized assays.

Keywords: Lyme disease; cholesterol; exome sequencing; machine learning.

This graphical abstract is also available at Tidbit https://tidbitapp.io/tidbits/burden-of-respiratory-viruses-in-children-less-than-two-years-in-a-community-based-longitudinal-birth-cohort

Figure 1.

Selection of study participants and rare predicted LoF and synonymous variants in cholesterol metabolism genes. The study cohort comprised 50 329 participants from the BioMe and 968 846 participants from the MSDW. A, Identification of cases of Lyme disease and controls. B, Ascertainment of rare LoF and synonymous variants in 5 cholesterol metabolism genes (APOB, APOE, CD36, LDLR, and PCSK9). LoF, splice acceptor or donor, stop gained, or frameshift consequence; synonymous, synonymous consequence; exclude, neither LoF nor synonymous (ie, inframe deletion or missense consequence). Abbreviations: AF, allele frequency; BioMe, BioMe Biobank; EHR, electronic health record; gnomAD, Genome Aggregation Database; Hx, history; LoF, loss of function; MANE, Matched Annotation from NCBI (National Center for Biotechnology Information) and EMBL-EBI (European Molecular Biology Laboratory-European Bioinformatics Institute); MSDW, Mount Sinai Data Warehouse; VEP, variant effect predictor.

Figure 2.

Total serum cholesterol, LDL-C, and genetically determined total cholesterol associate with risk of Lyme disease. Total serum cholesterol levels, serum LDL-C levels, cholesterol genetic scores, and cholesterol metabolism gene burdens were evaluated for association with Lyme disease in the BioMe Biobank. A, The proportions of individuals with clinically defined ranges of serum cholesterol (hypercholesterolemia, borderline, and desirable) and LDL-C (high, borderline, near optimal, and optimal) were determined for cases of Lyme disease and controls. Genetically determined cholesterol was analyzed for association with total serum cholesterol levels (B) and risk of Lyme disease (C). Standard deviation increases in total cholesterol and LDL-C genetic scores, burden of rare LoF variants in CD36 and LDLR, burden of rare LoF variants in APOB and PCSK9, and burden of rare synonymous variants in CD36 and LDLR (negative controls) were assessed with adjustment for age, sex, body mass index, and genetic ancestry. Adjusted change in total serum cholesterol (β) and adjusted OR for Lyme disease are depicted as points, and 95% CIs are shown as error bars. Diamonds depict meta-analyses with point estimates indicated by dashed vertical lines and 95% CIs extending the length of the diamond. Abbreviations: CI, confidence interval; LDL-C, low-density lipoprotein cholesterol; LoF, loss of function; OR, odds ratio.

Figure 3.

Profiling of total serum cholesterol and Igs over the disease course in cases of Lyme disease. Laboratory measurements of total serum cholesterol and IgA, IgG, and IgM were obtained, and the median was computed for 4 stages in time for each participant in the BioMe Biobank: baseline (>1 year before diagnosis), ≤1 year before diagnosis, ≤1 year after diagnosis, and >1 year after diagnosis. The date of diagnosis is indicated by the dashed vertical line, and the approximate acute stage of the disease is delineated by the gray box. The distributions of values are depicted as violin plots with the median and interquartile ranges overlaid as box-and-whisker plots. A, Distributions of values for carriers of rare LoF variants in CD36 and LDLR are shown as orange violin plots, and values for noncarriers are shown as blue violin plots. B, Distributions of values for individuals with high genetic scores (top quartile) and low genetic scores (bottom quartile) are shown as orange and blue violin plots, respectively. Abbreviations: Ig, immunoglobulin; LoF, loss of function.

Figure 4.

Performance of machine learning model using routine measurements of lipid panel, basic metabolic panel, and complete blood count to diagnose Lyme disease. Performance metrics are tabulated and illustrated with receiver-operating characteristic curves for the random forest-based machine learning model. The model was trained and validated in the MSDW cohort and tested in the BioMe cohort. Abbreviations: AUROC, area under the receiver operating characteristic curve; BioMe, BioMe Biobank; CI, confidence interval; MSDW, Mount Sinai Data Warehouse; NPV, negative predictive value; PPV, positive predictive value.